Alphabet’s Google (NASDAQ:GOOGL) (NASDAQ:GOOG) and OpenAI employees are raising their voices against the company’s plans to collaborate with the Pentagon on artificial intelligence technology.Over 100 Google staff members penned a letter to management expressing their concerns. They called for the establishment of clear boundaries or “red lines” in government contracts, akin to those pursued by Anthropic, reported the New York Times on Thursday.The employees took issue with the possible use of Google’s Gemi ...

Core Viewpoint - The article discusses the shift in energy supply dynamics in the U.S., particularly in relation to the growing demand for electricity from AI data centers and the challenges posed by aging infrastructure and extreme weather conditions. It highlights the response from tech giants and the government to address these issues while emphasizing the need for a more robust energy infrastructure. Group 1: U.S. Energy Supply Challenges - Extreme weather, aging infrastructure, and insufficient investment have led to frequent failures in U.S. power facilities, resulting in significant outages during winter [2] - The International Energy Agency (IEA) projects that electricity consumption by data centers in the U.S. will reach 183 TWh in 2024, accounting for approximately 4% of total electricity use, and is expected to double to 426 TWh by 2030, potentially exceeding 12% [2] - The capacity auction held by PJM, the largest regional grid operator in the U.S., saw prices reach $333.44 per megawatt, indicating that demand from data centers far exceeds new supply [3] Group 2: Government and Tech Giants' Response - The U.S. government is pushing tech giants to transition from being mere energy consumers to defining and owning energy infrastructure, with a focus on nuclear, gas turbine, and diversified storage technologies [5] - The Trump administration has set a goal to start construction on 10 new nuclear reactors by 2030, marking the largest nuclear power expansion in 30 years, with a strategic partnership worth $80 billion with Westinghouse Electric [6] - Google has announced significant partnerships for clean energy projects, including a 1.9 GW clean energy project utilizing a 100-hour long-duration storage solution, aiming to set a benchmark for the industry [6] Group 3: Market Implications and Opportunities - The energy infrastructure boom in the U.S. is creating opportunities for high-end manufacturing companies, including Chinese firms, as demand for gas turbines and energy storage equipment rises [7] - The U.S. AI data center energy market is projected to exceed $100 billion, presenting a significant market share opportunity for companies in the gas turbine and storage sectors [7] - China is also facing similar challenges with rising power demands due to AI developments, and is working on optimizing its electricity market mechanisms to better accommodate new load demands [8]

今天带大家扒一扒Contrary发布的《2026年科技趋势报告》这本352页的"未来预言书"。一句话总结：2026年，科技不再只是手机里的APP，而是从 天上到地下，从身体到情感，全方位重塑我们的生活，带来巨大的机遇和同样巨大的拉扯。 报告内容量巨大，我把它掰开揉碎，分成四大块，咱们用大白话唠明白。 读完整份报告，最大的感受是：未来已来，只是分布不均，而且充满悖论。 一方面，技术以一种近乎野蛮的方式生长。 AI重塑了我们的工作方式（从写代码到看病），自动驾驶和低空经济重塑了我们的物理空间（从陆地到 天空），生物技术甚至开始重塑我们的身体（从减肥到基因编辑）。这背后是中美两大巨头的激烈博弈，从芯片、能源到月球资源，每一个赛道都 关乎未来几十年的国运。 另一方面，技术进步与人类的精神需求产生了巨大的错位。 我们享受着屏幕带来的便利和娱乐，却陷入了更深层次的孤独和不信任。当现实中的社 交变得"昂贵"和困难时，我们甚至开始转向AI寻求情感慰藉。这给我们带来一个深刻的启示：未来最稀缺的资源，可能不再是算力、能源或者稀 土，而是真实的"人际关系"和"信任"。 对于普通人来说，这既是挑战也是机会。挑战在于，我们需要不断学习 ...

北京时间2月26日，三星推出旗舰款Galaxy S26系列手机。在谷歌Gemini模型的加持下，三星S26系列具 备了类似于豆包手机的自主订餐、打车等任务执行能力，但目前仅在美国和韩国上线了这项功能。 尽管三星手机国行版无法使用谷歌Gemini，但南都记者实测发现，S26系列国行版内置的Bixby助手提供 了"智能执行"模式。接收用户指令之后，它会模拟人类使用手机的方式，自动完成滴滴叫车、美团点外 卖、拼多多购物等常见任务。任务启动时，手机屏幕边缘会出现彩色光晕提示，和豆包手机的使用情形 相似。 三星方面未披露Bixby助手"智能执行"模式所接入的具体国产大模型。此前，三星和字节跳动、智谱开 展过合作，且这两家公司在手机智能体能力上均有代表性产品。其中，字节跳动旗下火山引擎曾在2025 年7月透露，三星Galaxy的Z系列产品引入了豆包大模型、火山引擎联网问答Agent（智能体）等产品， 新升级的Bixby助手等结合了豆包大模型的深度思考、多模态内容理解与生成能力。 针对Bixby助手"智能执行"模式是否有调用了旗下大模型，字节跳动和智谱方面均不作回应。 值得注意的是，谷歌并未完全依赖GUI方案，同时开发了一 ...

Core Insights - DeepMind's latest paper introduces AlphaEvolve, which treats algorithm source code as a genome and uses Gemini as a genetic operator to evolve new game theory algorithms through a process akin to natural selection [1][5][20] - The evolved algorithms outperform human-designed optimal solutions in various tests, utilizing mechanisms that human researchers had not previously considered [1][22] Group 1: AlphaEvolve and Its Mechanism - AlphaEvolve is described as an evolutionary coding agent that uses the source code as a genome, with LLM acting as a genetic operator to mutate the code [5][20] - The process involves evaluating the fitness of each "offspring algorithm" based on its exploitability in a set of benchmark games, allowing the best-performing algorithms to survive and evolve further [5][6][20] Group 2: Target Algorithms and Historical Context - The focus of AlphaEvolve is on two core algorithm families in multi-agent reinforcement learning: Counterfactual Regret Minimization (CFR) and Policy Space Response Oracles (PSRO) [6][7] - Historically, researchers have manually tuned and designed variations of these algorithms, but AlphaEvolve automates this process, significantly expanding the search space for potential solutions [10][16][20] Group 3: Implications of AI-Driven Algorithm Design - The paper highlights that the design of game theory algorithms has traditionally been a challenging task due to the complexity of incomplete information games [12][13] - AlphaEvolve's approach allows for meaningful mutations in code, leading to the discovery of effective strategies that human experts had not conceived [17][25] - The results indicate a paradigm shift where AI not only executes algorithms but also invents them, achieving superior performance compared to human-designed methods [22][25] Group 4: Future Directions - DeepMind plans to apply this evolutionary framework to the complete design of deep reinforcement learning agents and explore mechanism discovery in cooperative games [25]

传统AI模型在稀疏奖励环境中，往往会找不到激励难以学会层次化思考。如今，谷歌团队通过引入元控制器操控模型内部残 差流，让智能体学会了「跳跃式思考」。该研究揭示了大模型内部可自发形成了类似人脑的层次化决策机制，为AI在需要多 步的复杂任务提供了全新的训练范式。 AI智能体最大的「硬伤」，是算力不够？ 并不是，奖励太少、路太长才是。 在稀疏奖励的长序列任务里，传统token-by-token探索像蒙眼走迷宫：没有路标、没有提示，只有走到终点才知道对不对。 结果就是一个尴尬现实：想让智能体做点复杂事，往往必须外挂规划器「扶着走」。 而谷歌这项研究直接换打法：在迷宫里要求智能体按顺序踏过一串彩色子目标，且只有全程无误才给奖励——用最残酷的稀 疏奖励，逼出真正的层次化决策能力。 真正的突破在于：他们不再只优化输出，而是开始操控模型内部的「认知过程」。 在稀疏奖励下，智能体如何高效探索 传统的大模型，依赖逐词生成（token-by-token）的探索方式，而这对于需要多个正确步骤才能获得奖励的复杂任务，由于奖 励稀疏，导致智能体难以完成需要层次化决策的长序列任务。 这好比让一个人蒙着眼睛走迷宫，只有到达终点才能获得反馈， ...

（一）消费 1、彪马去年Q4营收15.6亿欧元，同比下降27.2% 瑞幸咖啡2月27日发布2025年财报，全年总净收入492.88亿元，同比增长43%，门店总数超3.1万家，累 计交易客户数突破4.5亿。 （二）出行 1、语音误关大灯致车祸 领克高管致歉 近日，一车主驾驶领克Z20夜间在高速上行驶时，语音操作误关大灯致车辆撞上护栏一事引发网友广泛 关注。针对此事，2月26日，穆军在其个人微博发文称："昨晚发生一起领克Z20车辆行驶中语音误操作 控制关闭大灯的情况，今天我们第一时间完成了语音控制优化方案，现已通过云端推送更新，后续在行 驶状态下只能通过手动控制大灯关闭，请大家放心。感谢用户的反馈与监督，对此带来的困扰我们深表 歉意，领克始终守护您的安全。" 2、吉利汽车：以约6312万港元回购384.1万股 2月26日，吉利汽车在港交所公告称，当日按每股16.31-16.66港元回购股份，加权均价16.43港元。当前 库存股达2345.6万股，占已发行总股本约0.22%。 3、梅赛德斯-奔驰中国与Momenta签署升级合作谅解备忘录 2月26日，彪马公布2025年Q4业绩，营收同比下降27.2%至15.6亿欧元 ...

Strip away the rhetoric, and the U.S. rare earth landscape collapses fast. Most companies are still upstream, with mines, oxides, separation pilots, and PowerPoint roadmaps. REalloys is firmly downstream, where supply chains either function or fail.That makes heavy rare earths absolutely vital to things like precision-guided missiles and missile-defense interceptors. Dysprosium and terbium are non-negotiable inputs for these weapons systems.Heavy rare earths keep modern missile and aerospace systems operati ...

①【三大期指齐跌】截至发稿，道指期货跌0.40%、标普500指数期货跌0.22%、纳指期货跌0.10%。 ②【中概股盘前涨跌不一】中概股盘前涨跌不一。截至发稿，阿里巴巴跌0.95%，拼多多涨0.41%，百 度涨0.88%，京东跌0.59%。 ③【英伟达反弹0.8%】继昨日英伟达收跌5.46%后，截至发稿，英伟达盘前反弹0.8%。消息面上，每日 经济新闻独家报道，全球最大的AI模型API聚合平台OpenRouter的数据显示，2026年2月，中国AI模型 的全球调用量实现了历史性的突破。在2月9日至15日当周，中国模型的周调用量达到4.12万亿Token， 首次超越了美国同期的2.94万亿Token。该消息引发了市场对于英伟达算力逻辑的担忧。中国大模型主 流MoE体系的出现使得Token使用量的爆发，不再与英伟达GPU的需求量线性挂钩。 ④【巨大资本开支引发市场担忧，CoreWeave跌近10%】云基础设施提供商CoreWeave盘前跌近10%。消 息面上，CoreWeave面临天量规模的积压订单，公司预计2026年资本支出为300亿至350亿美元，远高于 2025年的103.1亿美元。巨大的资本支出引发市场对 ...

谷歌趁热打铁，Nano Banana 2这就上线了！ 本来Nano Banana就还是AI生图届顶流，这一次2代版本推出，直接在竞技场上又超了Pro版本100分，主打一个我自己的王座自己来抢（doge）。 | = Rank by | 5 | | Text-to-Image Arena Text Rendering | | | | | --- | --- | --- | --- | --- | --- | --- | | | | | | View overall rankings across Al image models based on their ability to generate images that accurately deliver on text-based prompts. | | | | Models Labs | | ( Feb 25, 2026 | ▽ 992,257 votes 日 50 models | | | | | E Categories (8) | < | | | | | | | | | * Hide Filters | | | | Q | | Overall | | ...